Background

In a context of fast-growing population, more and more people are exposed to large devastating earthquakes (Bilham 2004; Jackson 2006). This is particularly true at the foot of the Himalayan range where such events are given to happen in the coming decades when, on the meantime, the population is expected to grow quickly and aggregate in supercities (Bilham 2009). Indeed, since the last major Himalayan earthquake, the giant M w ~8.6 August the 15th 1950 earthquake in Assam, the population of the whole Indian subcontinent more than tripled and concentrated in densely populated large cities having yearly growth rates in excess of 20 % (The Registrar General and Census Commissioner, India: censusindia.gov.in). With the recent occurrence of the deadly M w 7.8 Gorkha earthquake of April 25, 2015 (Adhikari et al. 2015), the estimation of damage and loss of life from future large earthquake becomes an even more pressing priority.

Among the fastest growing in the last decades and actually most dense areas are the Himalayan range foreland basins. These regions were particularly impacted by the 1897 Shillong (Oldham 1899), 1905 Kangra (Middlemiss 1910) and 1934 Bihar (Rana 1935; Dunn et al. 1939) and will be impacted by future earthquakes that will rupture inevitably the Main Himalayan Thrust along the foothills of the mountain range (Fig. 1), with a major seismic gap located between Dehli and Patna and focusing much of the current attention and debates (e.g., Rajendran et al. 2015; Pulla 2015).

Fig. 1
figure 1

The great Bihar–Nepal January 15, 1934, earthquake epicenter, red star (Chen and Molnar 1977), and the Gorkha April 25, 2016, epicenter, yellow star, in light of the present-day population density. The past and future great earthquakes ruptures correspond to the area in between the thick gray line with triangles corresponding to the Main Frontal Thrust and the white dotted line materializing approximately the position of the downdip end of the presently locked fault zone. The area delimited by the green polyline corresponds to the area affected by damages during the 1934 earthquakes (Isoseist VI MSK64 from this study). Red thick line with triangles corresponds to the minimum extension of the 1934 earthquake surface rupture (Sapkota et al. 2013) and red dotted line to minimum extension of the rupture at depth. The background image is a color-coded 2011 Landscan estimation of the population density

Among the past Himalayan earthquakes, the Bihar–Nepal, January 15, 1934, earthquake, with a death toll of more than 8000 people in Nepal and 7000 in India, deeply traumatized the population. In Nepal, several original testimonies were collected by Brahma Shumsher Rana, a Nepalese army “major general” responsible for the rescue and reconstruction operations (see Additional file 1). Some of these testimonies as well as key scientific information are compiled in his 1936 book, “Mahabukhampa”—“Great Earthquake” in nepali (Rana 1935). It has been partially complemented by the observations made during the three trips organized through the meizoseismal zone in eastern Nepal and Ganges basin (Dunn et al. 1939). A systematic macroseismic data collection was concomitantly organized in India. Indeed, the collection of questionnaires prepared by the Geological Survey of India and sent through the government of Bihar and Orissa were a major source of information in terms of macroseismic effects (e.g., Dunn et al. 1939). These macroseismic surveys resulted in a vast accumulation of data, including death toll and damages, consigned in Rana (1935) and Dunn et al. (1939), compiled and further analyzed in review studies (Pandey and Molnar 1988). However, information on the fatality rates was lacking due to limited availability of any individual and housing census at the time of the event. This lack limited exploitation of the data collected in terms of hazard estimation.

In this article, we determine the casualty rate and its spatial variations in Nepal. For that purpose, we confront the fatality counts with estimates of the Nepalese population repartition deduced from two National populations and housing census carried out in 1924 and 1942. We then extrapolate these fatalities to modern conditions. The difficulties of such extrapolations are illustrated by comparing the 1934 fatality rates with the fatality rates of the April 25, 2015, M w 7.8 Gorkha earthquake and also scaling laws proposed by Shiono (1995) for the Asian region.

Macroseismic dataset: from collection to interpretation

“It was exactly twenty-four minutes and twenty-two seconds after two PM Nepal time” on January the 15th 1934 (Magh 2 1990 in the Bikram Sambat calendar) “when a strange noise, assimilated to a rumble coming from the earth’s interior,” was perceived by the Kathmandu valley inhabitants. “This noise was followed by observations of water in reservoirs, basins and containers overflowing and spilling out. Observers then felt the ground moving from east to west before describing it as bended. The strong shaking followed immediately and its arrival induced development of cracks and the collapse of the first houses.” These observations described by Bhrama Shumsher Rana in 1935 (see Additional file 1 for his biography and the context of his study) are not limited to Kathmandu valley. Indeed, the author reported widespread damages from Kathmandu to the eastern Nepal border with India. The devastation affected also particularly northern India, and above all a large part of the state of Bihar. The strong damages reported were accompanied by a large number of casualties including a death toll greater than 15,000 people. Comprehensive macroseismic studies carried out in India (Dunn et al. 1939) had suggested at first that the epicenter was located in the Indian plains. Indeed, a 300-km-long region of Bihar, named the “slump belt,” was very strongly impacted, in soft ground area, by liquefactions and slumping as well as the place of a metrical subsidence measured by spirit leveling shortly after the earthquake [references and discussions in Bilham et al. (1998)].

B.S. Rana, from the beginning of his book (Rana 1935), refutes this thesis and gives the elements that make him think that the earthquake did happen in Nepal. Among others, he could actually see that many villages were destroyed east of Kathmandu, the damaged area far exceeding the state of Bihar and the Kathmandu valley. He also notes that many landslides were triggered in the east, in the vicinity of Udaypur Gadhi and Dharan (86–87.5°E). He further described the situation in the eastern mountains further north, near the village of Bhojpur (87° E), with the Sanskrit term “patala,” somewhat ambiguous in this context, but suggesting the idea of hell (Pandey and Molnar 1988). Concurrently to the first scenario, associating the epicentral region with the state of Bihar, the distribution of heavy destructions along the Nepal foothills as well as further north in the Lesser Himalaya of eastern Nepal rather suggested an epicenter on a fault further north between the front of the high Himalayan range and the Main Frontal Thrust (Rana 1935; Pandey and Molnar 1988).

This second scenario was first consolidated by the instrumental relocation of the epicenter 10 km south of Mount Everest (Chen and Molnar 1977) (Fig. 1). It was then definitely confirmed by the discovery of the traces of a 150-km-long surface rupture along the Main Frontal Thrust in eastern Nepal (Sapkota et al. 2013; Bollinger et al. 2014) (Fig. 1). In the meantime, Hough and Bilham (2008) proposed that the strong intensities in the Gangetic plains, in addition to local liquefaction and site effects, are mostly due to post-critical Moho reflection that led to the focus of an aggressive SmS seismic phase.

In addition to providing qualitative information on the earthquake rupture, the macroseismic data material collected in India and Nepal in 1934 is an invaluable source of quantified observations. The macroseismic questionnaire prepared by the Geological Survey in India has been adapted two times in 1934 to be able to collect the most comprehensive information possible (Dunn et al. 1939). Government reports, newspapers and other materials complemented this information. Besides its qualitative descriptions from Nepal, Rana (1935) also provided a detailed accounting of casualties and damage by location. These data were analyzed by Pandey and Molnar (1988) and seem accurate enough to be considered/investigated with some attention (see Table 1). Note, however, that the casualty counts in Nepal must be taken cautiously, given the number of remote villages that were hit by the earthquake and for which it is doubtful that reliable figures will ever be available. The document also provides statistics on the buildings completely destroyed, heavily cracked and slightly cracked (see Table 1).

Table 1 Fatalities count per district due to the January 15, 1934, earthquake (Rana 1935) and comparison with population estimates

Notwithstanding the great quality of this information, the spatial analysis of macroseismic observations is delicate. Indeed, these observations depict a significant local variability partially simulated by biases coming from the observer, the various source of macroseismic information as well as from natural variations inherent to the local geological conditions. These observations were further translated into various intensity scales including MMI (Dunn et al. 1939), MSK-64 (Ambraseys and Douglas 2004) and EMS-98 (Martin and Szeliga 2010) using some subjective choices. Except first-order differences coming from integration (Dunn et al. 1939) or depletion (Ambraseys and Douglas 2004; Martin and Szeliga 2010) of the effects of liquefactions of the soils on buildings, the differences between the interpretations are difficult to clarify [see Szeliga et al. 2010 for quantifications of differences between Ambraseys and Douglas (2004) and Martin and Szeliga (2010)]. Finally, the density of the macroseismic intensities being highly variable along strike the felt area, we tested several interpolation/kriging schemes in order to finally compare fatalities counts and intensities and avoid subjectivity.

Whatever the method used on the intensity dataset depleted from the observations at sites encompassing soil liquefactions, the surface covered by intensities greater than VIII, usually correlated along shallow dipping thrust with the fault segment that ruptured, is relatively stable on the order of 10,800 ± 1600 km2. It corresponds approximately to the trace of the 150-km-long surface rupture mapped in the field (Sapkota et al. 2013) (Fig. 2). In turn, the typical value kriged within a radius of 150–250 km of the epicenter is intensity VII, while intensity VI is reached within 200–300 km. This intensity decrease as a function of the epicentral distance is typical of the decrease predicted by the intensity attenuation laws calibrated with all Himalayan earthquakes (Ambraseys and Douglas 2004; Szeliga et al. 2010).

Fig. 2
figure 2

Macroseismic map of the 1934 earthquake. Isoseists interpolated from 806 MSK64 macroseismic intensities compiled in Ambraseys and Douglas (2004). Slump belt was severely affected by liquefactions and slumping (contour from Dunn et al. 1939). Red star is the instrumental epicenter of the main shock from Chen and Molnar (1977). Red polyline with triangles corresponds to the trace of the Main Frontal Thrust that ruptured in 1934 according to Sapkota et al. (2013). In gray, 1934 administrative districts borders and names. Refer to Table 1 for full district names

In order to avoid biases induced (1) by the variable quality of the sparse macroseismic data available in Nepal and (2) by the spatially unresolved accounting of the observations within a district associated with (3) an unresolved high variability of the geological and topographical site effects, we will further compare the destructions and fatalities to “estimated MSK intensities” deduced from Ambraseys and Douglas (2004) and associated with each district centroid. While more recent relations in EMS-98 intensities are presented in the literature for Himalayan earthquakes (Szeliga et al. 2010), we retain this relationship in MSK-64 because it facilitates direct comparison of our results with those of Shiono (1995).

Besides, the distribution of the collapsed buildings and fatalities are far from being as clearly related to the epicentral distance. Indeed, in the eastern districts to the south of the epicenter there are a considerable number of collapsed buildings (see Table 2), suggesting a strong effect of the earthquake despite a comparatively small number of victims. The effects diminish quickly eastward, with little damage to Darjeeling (Fig. 2). To the west, in the Kathmandu valley, the rate of victims per collapsed buildings appears higher than elsewhere suggesting a higher vulnerability of the population to the destructions, most probably due to the taller buildings and higher buildings density (multi-story masonry with mud cement) as well as local site conditions. These observations could therefore benefit from being confronted to demographical records.

Table 2 Fatality rate due to the January 15, 1934, earthquake in the Kathmandu valley

Fatality rate collection and analysis

Besides this information on the macroseismic field, we benefited from the individual and housing counts of the Nepalese population issued before (1921) and after the earthquake (1942) by the Central Bureau of Statistics, critical information that was not available to previous studies including Pandey and Molnar (1988). We estimated the population counts in 1934 from an extrapolation in between the values for these censuses, estimating the average annual growth rate for each region and making the hypothesis that these rates are constant over the considered period. Given that we did not benefit from details concerning the population repartition within the Kathmandu valley in 1921 and 1942, we used those from the 1953 census, making the hypothesis that the distribution of the population remained similar up to at least 1953 (Table 1). The growth rate remained smaller than 1 % per year during that period; therefore, uncertainties, conservatively taken as 5 % when necessary, have no consequence on the conclusions we reach in this study.

We then confront, to evaluate the fatality rate, the number of victims reported by Rana (1935) to this estimate of the 1934 population. These numbers are summarized in Tables 1 and 2. In Kathmandu valley, the fatality rate appears higher than in other areas of the meizoseismal zone in Nepal (Fig. 2). Indeed, the fatality rates in Kathmandu, Bhaktapur and Lalitpur–Patan districts are typically as high as 0.5 % or larger, while they rarely reach 0.1 % in most regions at similar distances from the epicenter. Actually, the fatality rate is of the order of 2 % for Patan, similar in and out the city center due to a mixed type of building, but varies in between 0.4 % for the most rural areas in Bhaktapur and Kathmandu to nearly 5 % for the urban community of Bhaktapur. The high fatality rate in that urban community is of the same order of magnitude as the fatality rates recorded in the worst fatal meizoseismal areas which include the cities of Managua, Nicaragua, in 1972 (1 %), Spitak, Armenia, in 1988 (4.5 %), Avezzano and Messina, Italy, in 1915 and 1908 (17 and 20 %) (e.g., Nichols and Beavers 2003, 2008) and has to be compared to the 30 % recorded in 1976 in areas of the city of Tangshan, China, exposed to total collapse of masonry buildings (Shiono 1995).

This high fatality rate in Kathmandu valley is probably due to combined effects of the vulnerable multi-storys building stock, made of brick with mud mortar, its high concentration and of additional seismo-geological effects typical of the Kathmandu basin, including (1) very long solicitation of the structures due to trapping of the seismic waves (e.g., Bhattarai et al. 2012; Chamlagain and Gautam 2015), (2) dominant seismic periods, related to the seismic source and sedimentary basin response, corresponding to the natural periods of multi-storys buildings (e.g., Paudyal et al. 2012; Rajaure et al. 2014; Goda et al. 2015; Galetzka et al. 2015; Bhattarai et al. 2015) and (3) liquefactions (e.g., Gajurel et al. 2000; Mugnier et al. 2011). In comparison, mountainous areas of Bhojpur and Udaypur Gadhi fatality rates, on the hanging wall of the fault that ruptured, are of the order of one percent, which is low compared with the values observed in Kathmandu valley, but still very high for rudimentary buildings, wood-framed with light roof material, usually safer when, in addition, established on the bedrock.

When presented as a function of the distance to epicenter (Fig. 3), the earthquake fatality rate decreases in a similar way as the macroseismic intensity, not taking into account the most densely populated areas of the Kathmandu valley. Indeed, those regions appear as outliers, far beyond average fatality rates, even when considering separately the most urban areas and their vicinity (Fig. 4). In 1934, Patan was a mix of urban and rural environment, whereas Bhaktapur urban and rural zones were (and actually still are) very pure, an observation that help in understanding the differences between the urban and rural end members of each city. However, these fatality rates from these urban centers appear better correlated when plotted relatively to the building destruction rate (Fig. 5), suggesting a possible causal relationship between the fatality rate and the population density.

Fig. 3
figure 3

Fatality rate from the 1934 Nepal–Bihar earthquake as a function of the distance to epicenter (Rana 1935). The dashed curve corresponds to the attenuation of the MSK macroseismic intensity from Ambraseys and Douglas (2004) as a function of the distance to epicenter. The black diamonds correspond to the samples within the Kathmandu basin

Fig. 4
figure 4

Fatality rate from the 1934 Nepal–Bihar earthquake (Rana 1935) as a function of the mean MSK intensity from Ambraseys and Douglas (2004). The black diamonds correspond to the samples within the Kathmandu basin. The dashed curve corresponds to Shiono (1995) standard—unreinforced masonry—fatality rate relationship as a function of the macroseismic intensity

Fig. 5
figure 5

Fatality rate from the 1934 Nepal–Bihar earthquake as a function of the building destruction (Rana 1935). The black diamonds correspond to the 1934 earthquake, the red diamonds to the 2015 Gorkha earthquake and the blue diamonds to the 2006 Muzaffarabad, Pakistan, earthquake (Maqsood and Schwarz 2011)

These fatality rates can already be used to derive first-order fatality estimates in the case of a repetition of the 1934 earthquake, assuming the local vulnerability has remained constant. For example, with the 2011 population census, we can calculate casualties in each zone used to calculate the 1934 fatality rates and reported in Table 1. Taking into account the change in administrative boundaries used in the recent census compared with 1934, we obtain numbers given in details in Additional file 1: Table S1. We then obtain 33,000 victims in Nepal, but over 50,000 in India. Due to the large increase in population in northern India, larger numbers of victims are expected in India, while the rupture zone remains in Nepal. If the 2001 census population is used instead, we obtain about 26,000 victims in Nepal and 39,000 in India. An interesting temporal variability can be observed when the 2001 and 2011 censuses are compared, beyond just the unavoidable increase in the number of victims. Population growth remains larger than 20 % per year in India, but is decreasing to values of the order of 14 % in Nepal. Hill districts of the epicentral zone of the 1934 earthquake tend to lose population, while Nepalese population dramatically increases in the foothills near the Indian border. Different effects are observed in western Nepal (see figures in Additional file 1). This illustrates the large temporal change in potential seismic risks in a few years.

The numbers of victims estimated above can only be considered as a baseline. Indeed, numerous effects complicate the problem. In addition to the population growth and redistribution, the fatality rates must have been modified by the change in population lifetime and the evolution of construction methods. The April 25, 2015, Gorkha earthquake definitely brings important new information to address these problems.

Comparison with the 2015 Gorkha earthquake

The earthquake of April 25, 2015 (Fig. 1), of magnitude M w ~7.8 (M L ~7.6), with more than 8700 victims, is the most deadly earthquake in Nepal since the M w ~8.4 megaquake of 1934. Beyond the large number of victims, the destruction of infrastructure and houses was tremendous in the villages north of Kathmandu, even total in some locations, and a terribly traumatic situation was created for the population. While the epicenter was located near Gorkha city, the aftershock distributions covered a 150-km segment extending to the east (Adhikari et al. 2015). Fatality rates were moderate in the Katmandu Valley (<0.1 %), but was surprisingly large in hill districts north of Kathmandu, with a maximum fatality of 1.5 % in Rasuwa District, 1.2 % in Sindhupalchowk district and 0.4 % in Nuwakot district (Central Department of Statistics, Home Ministry, Nepal). When comparing the fatality rate with the building destruction rate per inhabitant (Fig. 5), the fatalities of the Gorkha earthquake appear significantly smaller than for the 1934 earthquake, while the building destruction rate appears comparatively large for the 2015 Gorkha earthquake. Actually, in some rural communities north of Kathmandu, the building destruction rate was close to 100 % in 2015. This suggests that large differences in some rural and some urban districts have now emerged in Nepal, with the implementation of appreciably efficient building methods in Kathmandu valley, while the construction methods were not improved at all since the 1934 earthquake in rural communities. Note that the 1934 and 2015 earthquakes both happened on a holiday around noon time.

The intensity distribution in the case of the Gorkha earthquake (Martin et al. 2015) hardly compares with predictions from known attenuation laws of the macroseismic intensity at short distance (the first tenths of kilometers) from the source. Nevertheless, to compare with the previous analysis for the 1934 earthquake, we use again the Ambraseys and Douglas (2004) attenuation law to estimate the intensities, taking into account an effective macroseismic epicenter north of Kathmandu in the center of the aftershock distribution. We note that the macroseismic intensities predicted by this attenuation law are significantly larger than what was observed for the Kathmandu valley (Martin et al. 2015), 10 to 30 km from the ruptured fault segment (e.g., Avouac et al. 2015; Grandin et al. 2015; Kobayashi et al. 2015). The fatality rates versus the estimated intensities (Fig. 6) are significantly smaller than the values observed for the 1934 earthquake. This conclusion still hold after converting the estimated intensities in true intensities at observation sites documented both in 1934 and 2015, respectively, in Martin and Szeliga (2010) and Martin et al. (2015). In Fig. 6, scaling laws proposed by Shiono (1995) are also shown versus intensity. While the large heterogeneity mentioned before causes large scatter in the case of the Gorkha earthquake (e.g., intensities reported in Kathmandu valley typically range from I EMS98 6 to 8), the so-called composite building Shiono scaling law appears on average as a satisfactory description of the 2015 data.

Fig. 6
figure 6

Fatality rate from the 1934 Nepal–Bihar earthquake (Rana 1935) as a function of the mean MSK intensity estimated from Ambraseys and Douglas (2004). The black diamonds correspond to the 1934 earthquake and the red diamonds to the 2015 Gorkha earthquake. The full lines correspond to the scaling laws proposed by Shiono (1995). See text for discussions on the biases—significant for the Kathmandu valley—introduced by the “estimated MSK intensity” reference frame

Conclusions

In this paper, complementing previously known information, the censuses carried out in 1921 and 1942 in Nepal are used to evaluate the fatality rates from the great 1934 Bihar–Nepal earthquake and better characterize its impact in eastern Nepal. Such data provide important archives, given the scarcity of documented magnitude 8 earthquakes. While rarely exceeding 1 %, even in the most impacted areas of eastern Nepal, close to the epicenter and on the hanging wall of the thrust fault that was activated, the fatality rates exceed 0.5 % in Kathmandu valley, reaching even 5 % in urban Bhaktapur. Such observed fatality rates can be used to broadly estimate their order of magnitude in case this earthquake repeats. While the numbers obtained have to be taken with caution, they definitely illustrate that the country has to consider preparations. The order of magnitude of temporal variations (+15 % per decade) noted when using 2001 or 2011 population numbers, and the fact that larger numbers of victims have now to be expected on the India side, are probably robust.

Compared with the fatality rates of the 1934 earthquake, the fatalities of the 2015 earthquake are significantly smaller for a given estimated intensity, except in some hill districts. This suggests that the lessons of the 1934 earthquake may have been ignored in some rural districts and that earthquake prevention methods have not been widely implemented outside the Kathmandu valley. This is also illustrated by the fact that the composite Shiono scaling law (Shiono 1995), which reflects the building prevention strategies suggested in Asia after the 1976 Tangshan earthquake in China, appears to reproduce, on average, the 2015 fatality rates, while the scaling law for unreinforced masonry reproduces the 1934 earthquake. Therefore, while numerous problems remain open, the results of decades of earthquake prevention methods, indeed, seem to have saved a significant number of lives in 2015. In the absence of more precise work, the composite Shiono scaling law could be a reasonable approach to estimating casualties in the case of contemporary large Himalayan earthquakes.

While simple scaling laws, in no way, can be claimed as reliable predictive models, they suggest important features that can be used to guide further research. In any case, given the large increase in population crowding in buildings, which are built both in Nepal and in India without much consideration for earthquake hazards (e.g., Dixit et al. 2013), it is the duty of the scientific community to raise great concern in this matter, and to make sure that every possible steps are taken to mitigate the effects of the coming megaquake in the Himalayan foothills. Intensity and fatality distributions from the 2015 Gorkha earthquake, while referring to a smaller earthquake and probably of a different kind, will be of tremendous importance to better constrain possible fatality scaling laws, and the potential damage and loss of life from the next giant Himalayan earthquake.